Ghost Eye

Ghost Eye 👁️

Give your text-only LLM the power to see. Ghost Eye is a lightweight image preprocessing bridge: when an image enters the conversation, it calls any OpenAI-compatible vision model (default: nex-agi/Nex-N2-Pro), produces a structured plain-text output (full OCR + visual summary), and feeds it back into the conversation — so your pure-text model can "see" without ever touching a multimodal API.

What it does

User sends image → Ghost Eye detects → vision model analyzes → OCR text + scene summary → your LLM answers

Two modes

Mode 1: Auto-preprocess (recommended)

When multimodalPreprocess is configured, Ghost Eye fires automatically on any inbound image. The user never knows it's there — they just get answers about images.

Mode 2: Tool-call mode

Registered as analyze_image_by_nexn2 tool (name stays for backward compatibility). Your LLM calls it explicitly when it sees an image. Add this to the system prompt:

When the user sends an image, screenshot, photo, or document scan, call the analyze_image_by_nexn2 tool to extract text and describe the image, then answer based on the returned content.

Workflow

Step 1: Receive image

Priority: --image-path > --image-url > --image-base64

⛔ Always prefer --image-path to avoid command-line Argument list too long errors with large base64 strings. Only fall back to --image-url or --image-base64 when no local path is available.

# Preferred: local file path (no size limit)
python3 {baseDir}/scripts/analyze.py --image-path ""

# Fallback: public URL
python3 {baseDir}/scripts/analyze.py --image-url ""

# Last resort: base64 (small images only, <50KB)
python3 {baseDir}/scripts/analyze.py --image-base64 ""

The script handles:

• Format validation (JPG/PNG/WebP/GIF/BMP via magic bytes)
• Cache check / read / write (MD5-based, 7-day TTL)
• Image compression (Pillow, max 1920px longest edge, quality 85%)
• API call with 1 automatic retry
• Structured JSON output

Step 2: Parse JSON output

Success:

{"success": true, "content": "【OCR文字提取】\n...\n\n【画面内容总结】\n...", "metadata": {"model": "...", "tokens_used": 1200, "cached": false, "process_time_ms": 1500}}

Error:

{"success": false, "content": "error message", "metadata": {}}

Pass content directly into the LLM conversation context.

Step 3: Caching

• Cache directory: {baseDir}/cache/ (auto-created)
• Cache key: MD5 hash of raw image bytes
• TTL: 7 days (NEXN2_CACHE_TTL_DAYS)
• Clear cache: delete all .json files in {baseDir}/cache/
• Toggle: NEXN2_CACHE_ENABLE=true/false

Environment variables

If NEXN2_API_KEY is not set, returns a friendly Chinese error message.

Error handling

⚠️ Errors never crash the conversation — structured JSON response is always returned.

Setup

In openclaw.json under skills.entries:

"ghost-eye": {
  "enabled": true,
  "apiKey": { "source": "env", "provider": "default", "id": "NEXN2_API_KEY" },
  "env": {
    "NEXN2_BASE_URL": "https://api.siliconflow.cn/v1",
    "NEXN2_MODEL_NAME": "nex-agi/Nex-N2-Pro",
    "NEXN2_IMAGE_COMPRESS": "true",
    "NEXN2_CACHE_ENABLE": "true",
    "NEXN2_CACHE_TTL_DAYS": "7",
    "NEXN2_TIMEOUT_MS": "30000"
  }
}

For multimodalPreprocess auto-mode, see references/multimodal-config.md.

Supported platforms

• SiliconFlow (default, China-accessible)
• OpenRouter
• Any OpenAI-compatible chat completions endpoint

Safety

• ⛔ Image base64 is never logged or written to conversation text
• ⛔ API key is never hardcoded
• ⛔ Temporary files are cleaned up immediately after processing
• ⚠️ Output is always plain text / Markdown — never binary, never images